How Expensive Is Each Claude Model? (2026 Pricing Comparison)

Anthropic's general-access Claude API range currently has four main tiers:

Claude Sonnet 5's $2 input and $10 output pricing was originally introductory. Anthropic made it permanent on 10 August 2026, so the previously announced increase to $3 and $15 no longer applies.

Claude Mythos 5 has the same pricing and technical specifications as Fable 5, but access is limited to approved Project Glasswing customers rather than being available through normal self-service access.

The difference between the cheapest active model and the most expensive generally available model is substantial. Fable 5 costs ten times more than Haiku 4.5 for both input and output tokens.

Price per token, however, only tells part of the story. Claude can consume additional output tokens while reasoning, newer models use a tokenizer that produces more tokens for the same amount of text, and agentic workflows can repeatedly send conversation history and tool results back into the model. If you want capability rather than cost, see our best Claude model guide, the GPT-5.6 vs Claude Fable 5 deep dive, or the equivalent OpenAI pricing comparison.

How Claude API pricing works

Claude API usage is primarily billed through input and output tokens.

• Input tokens include instructions, conversation history, documents, images, tool definitions and tool results sent to the model.
• Output tokens include Claude's visible answer, tool calls and the tokens consumed during internal reasoning.

Claude's thinking tokens are charged at the model's output-token rate, including thinking tokens that are not displayed to the user. They also count toward the request's max_tokens limit.

Prompt caching introduces three additional rates:

A five-minute cache entry becomes less expensive after one successful reuse. A one-hour cache entry becomes less expensive after two successful reuses. This makes caching particularly valuable for long system prompts, product catalogues, policy documents, code repositories and conversation histories that remain stable across requests.

A realistic Claude model cost comparison

Prices per million tokens can be difficult to translate into an operational budget. The following comparison estimates the cost of processing 1,000 typical requests, where each request contains 2,000 uncached input tokens and 500 billable output tokens. Across 1,000 requests, this produces two million input tokens and 500,000 output tokens.

Estimated cost = (2 × input price) + (0.5 × output price)

The estimate includes any reasoning tokens within the assumed 500 output tokens. It does not include tools, cache writes, image inputs, data-residency premiums or additional agent turns.

Latest Claude model pricing

Anthropic's Batch API provides a 50 percent reduction on input and output token prices for asynchronous work.

Under the standard comparison:

• Haiku 4.5 costs half as much as Sonnet 5.
• Sonnet 5 currently costs 60 percent less than Opus 5.
• Sonnet 5 costs 60 percent less than Opus 5.
• Opus 5 costs half as much as Fable 5.
• Fable 5 costs ten times more than Haiku 4.5.

The pricing ladder is therefore relatively clean. Each step provides access to greater capability, but using the highest tier for every request can multiply the model bill without improving routine tasks.

Claude Haiku 4.5 pricing

Claude Haiku 4.5 costs:

• $1 per million input tokens
• $5 per million output tokens
• $1.25 per million tokens for a five-minute cache write
• $2 per million tokens for a one-hour cache write
• $0.10 per million cached input tokens
• $0.50 input and $2.50 output through the Batch API

It has a 200,000-token context window and supports outputs of up to 64,000 tokens. Anthropic describes it as the fastest current Claude model.

Haiku is the strongest option for inexpensive, high-volume work such as classification, data extraction, reply categorisation, routing, summarisation and customer-support triage.

At the example workload, one million requests would cost approximately $4,500 at standard rates or $2,250 through the Batch API.

The main limitation is capability rather than context length. A cheap model can become expensive operationally when its errors cause retries, manual review or failures in downstream workflows. Haiku should therefore be evaluated on a representative sample before it is used for decisions with significant commercial consequences.

Claude Sonnet 5 pricing

Claude Sonnet 5 currently costs:

• $2 per million input tokens
• $10 per million output tokens
• $2.50 per million tokens for a five-minute cache write
• $4 per million tokens for a one-hour cache write
• $0.20 per million cached input tokens

Anthropic confirmed on 10 August 2026 that these rates are permanent. The previously scheduled increase to $3 input and $15 output was cancelled.

Its Batch API price remains $1 input and $5 output.

Sonnet 5 has a one-million-token context window, supports up to 128,000 output tokens and uses adaptive thinking by default. Anthropic positions it as the best combination of speed and intelligence.

It is the most practical default for business applications, coding systems, research tools, document processing and operational agents.

The permanent price makes Sonnet 5 especially competitive. In the comparison scenario, it costs $9 per 1,000 requests, compared with $4.50 for Haiku and $22.50 for Opus. Cost forecasts can now use these rates without planning for an increase.

Claude Opus 5 pricing

Claude Opus 5 costs:

• $5 per million input tokens
• $25 per million output tokens
• $6.25 per million tokens for a five-minute cache write
• $10 per million tokens for a one-hour cache write
• $0.50 per million cached input tokens
• $2.50 input and $12.50 output through the Batch API

It has a one-million-token context window, supports 128,000 output tokens and uses adaptive thinking by default. Anthropic recommends it for complex agentic coding and enterprise work.

The example workload costs $22.50 per 1,000 requests. That is 2.5 times the current Sonnet 5 cost and five times the Haiku 4.5 cost.

Opus 5 is appropriate for difficult software engineering, multi-stage agents, financial or professional analysis, complex research and tasks where verification and judgment materially affect the outcome.

Anthropic launched Opus 5 at the same token price as Opus 4.8. This means an existing Opus 4.8 deployment does not receive a direct token-price saving by staying on the older model. Migration should still be evaluated for behavioural and compatibility differences, but the older model has no headline price advantage.

Claude Fable 5.1 pricing

Claude Fable 5.1, released on 1 September 2026, is now Anthropic's most capable generally available model. It costs:

• $10 per million input tokens
• $50 per million output tokens
• $12.50 per million tokens for a five-minute cache write
• $20 per million tokens for a one-hour cache write
• $0.25 per million cached input tokens
• $5 input and $25 output through the Batch API

Headline input and output rates are unchanged from Fable 5, but the cache read price has been cut to a quarter of the previous $1 rate. Anthropic estimates that reduces typical token-billed costs by around 25%, and by up to roughly 45% for highly agentic workloads that repeatedly re-read the same context. If you run long agent loops, that single change is the main reason to migrate.

It keeps the one-million-token context window, 128,000-token maximum output and always-on adaptive thinking, and adds a June 2026 knowledge cutoff. Claude Mythos 5.1 shares its specifications and pricing but remains invitation-only through Project Glasswing.

Because the uncached rates are identical, the example workload still costs $45 per 1,000 requests without caching. The saving appears once prompt caching is in use, which is exactly where long-context agents spend most of their budget.

Claude Fable 5 pricing

Claude Fable 5 costs:

• $10 per million input tokens
• $50 per million output tokens
• $12.50 per million tokens for a five-minute cache write
• $20 per million tokens for a one-hour cache write
• $1 per million cached input tokens
• $5 input and $25 output through the Batch API

It has a one-million-token context window, supports 128,000 output tokens and keeps adaptive thinking permanently enabled.

The example workload costs $45 per 1,000 requests. That is twice the price of Opus 5 and ten times the price of Haiku 4.5.

Fable should be reserved for the most demanding work, including long-running autonomous agents, difficult scientific or professional research and tasks that consistently fail on Opus.

Fable 5 has been superseded by Fable 5.1, which matches its input and output rates while charging far less for cache reads, so there is no reason to start a new deployment on Fable 5. Opus 5 is positioned as approaching Fable-level intelligence at half the token price, making Opus the more economical premium default for most organisations.

Claude Mythos 5 pricing

Claude Mythos 5 has the same $10 input and $50 output pricing as Fable 5.

It also shares Fable's one-million-token context window, 128,000-token output limit and always-on adaptive thinking. Access is limited to approved defensive cybersecurity and infrastructure customers through Project Glasswing. There is no standard self-service sign-up.

Mythos therefore does not belong in a normal commercial model-routing strategy. Its inclusion in the pricing table is relevant for completeness, but it is a specialist-access model rather than a general alternative to Fable or Opus.

How expensive are the older Claude models?

Anthropic continues to publish pricing for active previous-generation models and some models that remain available through partner-operated cloud platforms.

These statuses apply to Anthropic-operated platforms. Bedrock and Google Cloud can follow different retirement schedules.

The most important pricing conclusions are straightforward.

• Claude Opus 4.5 through Opus 4.8 cost exactly the same per token as Opus 5. Staying on one of these older models provides no standard API token saving.
• Claude Opus 4.1 is three times more expensive than Opus 5 on both input and output. It is also scheduled to retire from Anthropic-operated platforms on 5 August 2026. An active Opus 4.1 deployment should be treated as an urgent migration project.
• Claude Sonnet 4.6 and 4.5 cost $3 input and $15 output. Sonnet 5 is one-third cheaper per token than either older Sonnet model, permanently.
• Claude Haiku 3.5 has a lower listed rate than Haiku 4.5, but it is retired on Anthropic-operated platforms. Its price does not make it a suitable choice for a new first-party deployment.

The tokenizer changes the real cost

Claude 4.7 and later models use a newer tokenizer. Anthropic states that the same text produces approximately 30 percent more tokens than it does on earlier models, although the precise change depends on the content and workload.

This means a fixed-token chart does not fully represent the cost of processing the same document across model generations.

For example, Sonnet 5's current rate is two-thirds of Sonnet 4.6's rate. Applying a rough 30 percent token increase gives:

0.667 × 1.30 = 0.867

Under that simplified assumption, Sonnet 5's permanent rate could make it approximately 13 percent less expensive than Sonnet 4.6 for an equivalent amount of text.

This is a planning estimate rather than a guaranteed outcome. Output length, reasoning depth, tool usage and model behaviour can change substantially between versions. Anthropic recommends recounting prompts against Sonnet 5 rather than reusing token counts measured on Sonnet 4.6.

Claude's one-million-token context window has no long-context premium

Claude 4.6 and later models include the full one-million-token context window at standard token rates. A request containing 900,000 input tokens has the same per-token rate as one containing 9,000 tokens.

Prompt-caching and Batch API discounts also continue to apply across the full context window.

This is commercially significant for long documents, code repositories, financial records and long-running agents. A large context can still create a large bill because more tokens are being processed, but the price per token does not increase after a long-context threshold.

A 900,000-token uncached input would cost approximately:

Repeatedly sending the same 900,000-token context would be inefficient. Prompt caching can reduce subsequent cache-hit charges to one tenth of the normal input rate.

Batch processing can halve the bill

Anthropic's Batch API provides a 50 percent discount on both input and output tokens. That makes it suitable for:

• Lead and data enrichment
• Document extraction
• Classification
• Evaluations
• Bulk summarisation
• Overnight processing
• Large back-office jobs without an immediate response requirement

A workload costing $9 on Sonnet 5 at standard rates would cost approximately $4.50 through Batch. A $45 Fable workload would fall to $22.50.

For high-volume automation, deciding whether the result must be synchronous can have as much effect on the bill as selecting a cheaper model.

Fast mode doubles Opus token prices

Fast mode is available for Claude Opus 5 and Opus 4.8 through Anthropic's first-party API. It increases the rate to:

• $10 per million input tokens
• $50 per million output tokens

Anthropic states that fast mode can provide up to approximately 2.5 times higher output speed. It is unavailable through Batch and unavailable on partner-operated cloud platforms.

The example Opus workload therefore increases from $22.50 to $45.

Fast mode should be used when latency has direct commercial value, such as an interactive coding environment, live professional assistant or time-sensitive customer experience. It is difficult to justify for asynchronous or internal back-office work.

Additional costs that can change the bill

US-only inference applies a ten percent multiplier to Claude 4.6 and later models. The multiplier applies to input, output, cache writes and cache hits.

Web search costs $10 per 1,000 searches in addition to the tokens consumed by search results and the final model response. Web fetch has no separate tool fee, but the fetched content enters the context and is charged at the selected model's input rate.

Claude Managed Agents add a runtime charge of $0.08 per running session-hour, alongside the normal model token costs. Batch, fast-mode and data-residency modifiers do not apply to Managed Agents sessions.

How expensive are image inputs?

Current Claude models can process images, but images are converted into visual tokens and billed using the model's normal input-token rate.

A 1,000 by 1,000 pixel image uses approximately 1,296 visual tokens. Anthropic estimates that processing 1,000 such images costs approximately:

Models from Claude 4.7 onward support a higher image-resolution tier, which can consume more visual tokens. A high-resolution 4K image can use up to 4,784 visual tokens.

For screenshot agents, document vision and computer-use systems, image resolution can therefore be an important cost-control setting.

Which Claude model should you use?

Anthropic's own general guidance is to use Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning. Its model overview recommends Opus 5 for complex agentic coding and enterprise work, with Fable 5 reserved for workloads requiring the highest available capability.

Best low-cost Claude model: Haiku 4.5

Haiku 4.5 is the appropriate starting point for predictable and high-volume tasks. Typical uses include classification, extraction, categorisation, routing, basic summarisation and customer-support triage. It costs $4.50 per 1,000 requests in the comparison, or $2.25 through Batch.

Best general Claude model: Sonnet 5

Sonnet 5 is the strongest default for most new production applications. It provides a one-million-token context window, fast comparative latency, adaptive thinking and stronger agentic capability than previous Sonnet models. Its $2 and $10 rates became permanent on 10 August 2026, so new budgets can plan around them.

Best premium Claude model: Opus 5

Opus 5 should handle difficult tasks that require stronger reasoning, autonomy, verification or professional judgment. It is the best premium choice for complex coding agents, difficult research and high-value enterprise workflows. Its price is half Fable 5's rate, while Anthropic describes it as approaching Fable capability on a range of valuable tasks.

Highest-capability Claude model: Fable 5.1

Fable 5.1 is appropriate where the task's value outweighs the additional cost. It should generally sit behind an escalation rule rather than process every request. A system can first attempt the task with Sonnet or Opus and route only the most difficult failures to Fable.

Specialist-access model: Mythos 5

Mythos 5 is designed for approved defensive cybersecurity and infrastructure work. It is not a normal alternative for general commercial applications.

The recommended production setup

A routed model architecture is more cost-effective than selecting one model for every request.

The most economical structure uses Haiku for repetitive operational work, Sonnet as the normal production model, Opus for difficult exceptions and Fable only for requests where the additional capability has measurable value.

Batch processing should handle non-urgent bulk jobs, while prompt caching should be enabled when long instructions or documents are repeatedly reused.

Final recommendation

• Claude Sonnet 5 is the best overall default for most new Claude API applications, and its $2 and $10 pricing is now permanent.
• Claude Haiku 4.5 is the best low-cost active model. It should handle high-volume, predictable work where evaluations demonstrate an acceptable error rate.
• Claude Opus 5 is the best premium model. It provides a clearer value proposition than previous Opus versions because Opus 4.5 through 4.8 have the same token price.
• Claude Fable 5.1 replaces Fable 5 at the same input and output rates with cache reads at $0.25 per million tokens. It should still be used selectively: It costs twice as much as Opus 5 and ten times as much as Haiku 4.5 in the comparison scenario.
• Older models should generally be migration candidates rather than new selections. Opus 4.1 requires particularly urgent attention because it costs three times more than Opus 5 and is scheduled to retire on 5 August 2026.

Want the same cost discipline applied to your own stack? Compare capabilities in our AI model comparison tool, or see how we design and build model-routed automation systems.